> DS has been overly concerned about cost control in the past yr & how it can serve models cheaply. that's a shallow reading They barely cared about product. Until V4 their API was unusably slow. They want to be able to ignore bottlenecks in scaling RL at very long contexts.
Compare GLM-5.2 to DeepSeek V4-Pro: DS has been overly concerned about cost control in the past yr & how it can serve models cheaply. Zai has focused on developing the best model for SWE & users clearly don't mind spending extra for that.
If the Zai Aug release is another big leap like GLM-5.2, I will also pay for a more expensive plan.
Even their Max plan really isn't that expensive if it brings you top end performance.
this is the thing driving me crazy about DeepSeek. it spent the entire 2 yrs concerned about being efficient & not having to fund raise. Every new release, it's more about how cheap their model is rather than how much better it is for customers to use it.
if you try GLM-5.2 on ZCode, it's an entirely different experience than V4 or even K2.7 with Kimi Code (different in a good way).